MiniCPM-2B is an edge large language model jointly developed by ModelBest and the Tsinghua NLP Lab, with 2.4B parameters excluding embeddings.
Technical highlights
：After SFT, the comprehensive performance surpasses Llama2-13B, MPT-30B, Falcon-40B and other larger parameter models, and the Chinese, mathematics, and coding capabilities are better than Mistral-7B.
：After DPO alignment, it surpassed mainstream conversation models such as Llama2-70B-Chat and Vicuna-33B in MTBench evaluation.
：MiniCPM-V achieves the best performance among multi-modal models of the same scale, and some tasks are comparable to Qwen-VL-Chat with 9.6B parameters.
：MiniCPM supports deployment and inference on smartphones, and the streaming output speed exceeds human speech speed.
：MiniCPM-V is the industry's first multi-modal large model that can run on smartphones, breaking through the bottleneck of end-side multi-modal applications.
：Parameter Efficient Fine-Tuning (PEFT) requires only a single 1080/2080 level GPU.
：Full parameter fine-tuning requires only a single 3090/4090 level GPU, significantly reducing the hardware cost of model customization and iteration.